File size: 41,175 Bytes
b847d03
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
---
tags:
- fp8
- vllm
---

# Phi-3-mini-128k-instruct-FP8

## Model Overview
* <h3 style="display: inline;">Model Architecture:</h3> Based on and identical to the Phi-3-mini-128k-instruct-FP8 architecture
* <h3 style="display: inline;">Model Optimizations:</h3> Weights and activations quantized to FP8
* <h3 style="display: inline;">Release Date:</h3> June 29, 2024
* <h3 style="display: inline;">Model Developers:</h3> Neural Magic

Phi-3-mini-128k-instruct-FP8 quantized to FP8 weights and activations using per-tensor quantization through the [AutoFP8 repository](https://github.com/neuralmagic/AutoFP8), ready for inference with vLLM >= 0.5.0. 
Calibrated with 10 repeats of each token in the tokenizer in random order to achieve 99% performance recovery on the Open LLM Benchmark evaluations. 
Reduces space on disk by ~50%.
Part of the [FP8 LLMs for vLLM collection](https://huggingface.co/collections/neuralmagic/fp8-llms-for-vllm-666742ed2b78b7ac8df13127).


## Usage and Creation
Produced using AutoFP8 with random tokens as calibration, based on [AutoFP8 with calibration samples from ultrachat](https://github.com/neuralmagic/AutoFP8/blob/147fa4d9e1a90ef8a93f96fc7d9c33056ddc017a/example_dataset.py).

```python
from datasets import load_dataset
from transformers import AutoTokenizer
import numpy as np
import torch

from auto_fp8 import AutoFP8ForCausalLM, BaseQuantizeConfig

MODEL_DIR = "microsoft/Phi-3-mini-128k-instruct"
final_model_dir = MODEL_DIR.split("/")[-1]

CONTEXT_LENGTH = 4096
NUM_SAMPLES = 512
NUM_REPEATS = 10

pretrained_model_dir = MODEL_DIR
tokenizer = AutoTokenizer.from_pretrained(pretrained_model_dir, use_fast=True, model_max_length=CONTEXT_LENGTH)
tokenizer.pad_token = tokenizer.eos_token

tokenizer_num_tokens = len(list(tokenizer.get_vocab().values()))
total_token_samples = NUM_REPEATS * tokenizer_num_tokens
num_random_samp = -(-total_token_samples // CONTEXT_LENGTH)

input_ids = np.tile(np.arange(tokenizer_num_tokens), NUM_REPEATS + 1)[:num_random_samp * CONTEXT_LENGTH]
np.random.shuffle(input_ids)
input_ids = input_ids.reshape(num_random_samp, CONTEXT_LENGTH)
input_ids = torch.tensor(input_ids, dtype=torch.int64).to("cuda")

quantize_config = BaseQuantizeConfig(
    quant_method="fp8",
    activation_scheme="static",
)

examples = input_ids

model = AutoFP8ForCausalLM.from_pretrained(pretrained_model_dir, quantize_config=quantize_config)

model.quantize(examples)

quantized_model_dir = f"{final_model_dir}-FP8"
model.save_quantized(quantized_model_dir)
```

Evaluated through a modified version of vLLM with the following script:

```
#!/bin/bash

# Example usage:
# CUDA_VISIBLE_DEVICES=0 ./eval_openllm.sh "neuralmagic/Llama-2-7b-chat-hf-FP8" "tensor_parallel_size=1,max_model_len=4096,add_bos_token=True,gpu_memory_utilization=0.7"

export MODEL_DIR=${1}
export MODEL_ARGS=${2}

declare -A tasks_fewshot=(
    ["arc_challenge"]=25
    ["winogrande"]=5
    ["truthfulqa_mc2"]=0
    ["hellaswag"]=10
    ["mmlu"]=5
    ["gsm8k"]=5
)

declare -A batch_sizes=(
    ["arc_challenge"]="auto"
    ["winogrande"]="auto"
    ["truthfulqa_mc2"]="auto"
    ["hellaswag"]="auto"
    ["mmlu"]=1
    ["gsm8k"]="auto"
)

for TASK in "${!tasks_fewshot[@]}"; do
    NUM_FEWSHOT=${tasks_fewshot[$TASK]}
    BATCH_SIZE=${batch_sizes[$TASK]}
    lm_eval --model vllm \
        --model_args pretrained=$MODEL_DIR,$MODEL_ARGS \
        --tasks ${TASK} \
        --num_fewshot ${NUM_FEWSHOT} \
        --write_out \
        --show_config \
        --device cuda \
        --batch_size ${BATCH_SIZE} \
        --output_path="results/${TASK}"
done
```

In vllm=0.5.0, Phi-3 models are not fully supported, and running the above script will yield an AssertionError. However, replacing the file that throws an error with the file below will fix the issue. 


```
from abc import abstractmethod
from typing import Dict, List, Optional, Tuple

import torch
import torch.nn.functional as F
from torch.nn.parameter import Parameter

from vllm.distributed import (divide, get_tensor_model_parallel_rank,
                              get_tensor_model_parallel_world_size,
                              split_tensor_along_last_dim,
                              tensor_model_parallel_all_gather,
                              tensor_model_parallel_all_reduce)
from vllm.logger import init_logger
from vllm.model_executor.layers.quantization.base_config import (
    QuantizationConfig, QuantizeMethodBase)
from vllm.model_executor.utils import set_weight_attrs

logger = init_logger(__name__)


def adjust_marlin_shard(param, shard_size, shard_offset):
    marlin_tile_size = getattr(param, "marlin_tile_size", None)
    if marlin_tile_size is None:
        return shard_size, shard_offset

    return shard_size * marlin_tile_size, shard_offset * marlin_tile_size


def adjust_bitsandbytes_shard(param: Parameter,
                              qkv_offsets: Dict[str, Tuple[int, int]],
                              loaded_shard_id: str) -> Tuple[int, int]:
    """Adjust the quantization offsets and sizes for BitsAndBytes sharding."""

    total, _ = qkv_offsets["total"]
    orig_offset, orig_size = qkv_offsets[loaded_shard_id]

    quantized_total = param.data.shape[0]
    quantized_offset = orig_offset * quantized_total // total
    quantized_size = orig_size * quantized_total // total

    return quantized_size, quantized_offset


class LinearMethodBase(QuantizeMethodBase):
    """Base class for different (maybe quantized) linear methods."""

    @abstractmethod
    def create_weights(self, layer: torch.nn.Module,
                       input_size_per_partition: int,
                       output_partition_sizes: List[int], input_size: int,
                       output_size: int, params_dtype: torch.dtype,
                       **extra_weight_attrs):
        """Create weights for a linear layer. 
           The weights will be set as attributes of the layer.

        Args:
            layer: The layer that is using the LinearMethodBase factory.
            input_size_per_partition: Size of the weight input dim on rank X.
            output_partition_sizes: Sizes of the output dim of each logical 
                weight on rank X. E.g., output_partition_sizes for QKVLinear
                is a list contains the width of Wq, Wk, Wv on rank X.
            input_size: Size of the input dim of the weight across all ranks.
            output_size: Size of the output dim of the weight across all ranks.
            params_dtype: Datatype of the parameters.
        """
        raise NotImplementedError

    @abstractmethod
    def apply(self,
              layer: torch.nn.Module,
              x: torch.Tensor,
              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
        """Apply the weights in layer to the input tensor.
        Expects create_weights to have been called before on the layer."""
        raise NotImplementedError


class UnquantizedLinearMethod(LinearMethodBase):
    """Linear method without quantization.

    Args:
        separate_bias_add: If true, add bias separately after matrix
                           multiplication.
    """

    def __init__(self, separate_bias_add: bool = False):
        self.separate_bias_add = separate_bias_add

    def create_weights(self, layer: torch.nn.Module,
                       input_size_per_partition: int,
                       output_partition_sizes: List[int], input_size: int,
                       output_size: int, params_dtype: torch.dtype,
                       **extra_weight_attrs):
        weight = Parameter(torch.empty(sum(output_partition_sizes),
                                       input_size_per_partition,
                                       dtype=params_dtype),
                           requires_grad=False)
        set_weight_attrs(weight, {"input_dim": 1, "output_dim": 0})
        layer.register_parameter("weight", weight)
        set_weight_attrs(weight, extra_weight_attrs)

    def apply(self,
              layer: torch.nn.Module,
              x: torch.Tensor,
              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
        weight = layer.weight
        if self.separate_bias_add:
            if bias is not None:
                return F.linear(x, weight) + bias
            return F.linear(x, weight)
        return F.linear(x, weight, bias)


class LinearBase(torch.nn.Module):
    """Base linear layer.

    Args:
        input_size: input dimension of the linear layer.
        output_size: output dimension of the linear layer.
        bias: If true, add bias.
        skip_bias_add: If true, skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
        quant_config: Quantization configure.
    """

    def __init__(
        self,
        input_size: int,
        output_size: int,
        skip_bias_add: bool = False,
        params_dtype: Optional[torch.dtype] = None,
        quant_config: Optional[QuantizationConfig] = None,
    ):
        super().__init__()

        # Keep input parameters
        self.input_size = input_size
        self.output_size = output_size
        self.skip_bias_add = skip_bias_add
        if params_dtype is None:
            params_dtype = torch.get_default_dtype()
        self.params_dtype = params_dtype
        if quant_config is None:
            self.quant_method: Optional[
                QuantizeMethodBase] = UnquantizedLinearMethod()
        else:
            self.quant_method = quant_config.get_quant_method(self)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        raise NotImplementedError


class ReplicatedLinear(LinearBase):
    """Replicated linear layer.

    Args:
        input_size: input dimension of the linear layer.
        output_size: output dimension of the linear layer.
        bias: If true, add bias.
        skip_bias_add: If true, skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
        quant_config: Quantization configure.
    """

    def __init__(self,
                 input_size: int,
                 output_size: int,
                 bias: bool = True,
                 skip_bias_add: bool = False,
                 params_dtype: Optional[torch.dtype] = None,
                 quant_config: Optional[QuantizationConfig] = None):
        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
                         quant_config)

        # All the linear layer supports quant method.
        assert self.quant_method is not None
        self.quant_method.create_weights(self, self.input_size,
                                         [self.output_size], self.input_size,
                                         self.output_size, self.params_dtype)

        if bias:
            self.bias = Parameter(
                torch.empty(self.output_size, dtype=self.params_dtype))
            set_weight_attrs(self.bias, {"output_dim": 0})
        else:
            self.register_parameter("bias", None)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        bias = self.bias if not self.skip_bias_add else None
        assert self.quant_method is not None
        output = self.quant_method.apply(self, x, bias)
        output_bias = self.bias if self.skip_bias_add else None
        return output, output_bias

    def extra_repr(self) -> str:
        s = f"in_features={self.input_size}"
        s += f", output_features={self.output_size}"
        s += f", bias={self.bias is not None}"
        return s


class ColumnParallelLinear(LinearBase):
    """Linear layer with column parallelism.

    The linear layer is defined as Y = XA + b. A is parallelized along
    its second dimension as A = [A_1, ..., A_p].

    Args:
        input_size: first dimension of matrix A.
        output_size: second dimension of matrix A.
        bias: If true, add bias.
        gather_output: If true, call all-gather on output and make Y available
                       to all GPUs, otherwise, every GPU will have its output
                       which is Y_i = XA_i
        skip_bias_add: This was added to enable performance optimizations where
                       bias can be fused with other element-wise operations. we
                       skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
        quant_config: Quantization configure.
        output_sizes: list of output sizes packed into one output, like for QKV
                       the list would be size 3.
    """

    def __init__(self,
                 input_size: int,
                 output_size: int,
                 bias: bool = True,
                 gather_output: bool = False,
                 skip_bias_add: bool = False,
                 params_dtype: Optional[torch.dtype] = None,
                 quant_config: Optional[QuantizationConfig] = None,
                 output_sizes: Optional[List[int]] = None):
        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
                         quant_config)

        self.gather_output = gather_output

        # Divide the weight matrix along the last dimension.
        tp_size = get_tensor_model_parallel_world_size()
        assert self.quant_method is not None
        self.output_size_per_partition = divide(self.output_size, tp_size)
        self.output_partition_sizes = [self.output_size_per_partition]
        # If QKV or MergedColumn, use output size of each partition.
        if hasattr(self, "output_sizes"):
            self.output_partition_sizes = [
                divide(output_size, tp_size)
                for output_size in self.output_sizes
            ]

        if output_sizes is None:
            output_sizes = [output_size]
        self.quant_method.create_weights(
            layer=self,
            input_size_per_partition=self.input_size,
            output_partition_sizes=self.output_partition_sizes,
            input_size=self.input_size,
            output_size=self.output_size,
            params_dtype=self.params_dtype,
            weight_loader=self.weight_loader)
        if bias:
            self.bias = Parameter(
                torch.empty(self.output_size_per_partition,
                            dtype=params_dtype))
            set_weight_attrs(self.bias, {
                "output_dim": 0,
                "weight_loader": self.weight_loader,
            })
        else:
            self.register_parameter("bias", None)

    def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
        # Special case for Fp8 scales.
        fp8_scales_shard_indexer = getattr(param, "fp8_scales_shard_indexer",
                                           None)

        tp_rank = get_tensor_model_parallel_rank()
        output_dim = getattr(param, "output_dim", None)
        param_data = param.data
        if output_dim is not None:
            shard_size = param_data.shape[output_dim]
            start_idx = tp_rank * shard_size
            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                 shard_size)
        # Special case for Fp8 scales.
        elif fp8_scales_shard_indexer is not None:
            param_data, loaded_weight = fp8_scales_shard_indexer(param_data,
                                                                 loaded_weight,
                                                                 shard_id=0)

        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)

    def forward(self, input_):
        bias = self.bias if not self.skip_bias_add else None

        # Matrix multiply.
        assert self.quant_method is not None
        output_parallel = self.quant_method.apply(self, input_, bias)
        if self.gather_output:
            # All-gather across the partitions.
            output = tensor_model_parallel_all_gather(output_parallel)
        else:
            output = output_parallel
        output_bias = self.bias if self.skip_bias_add else None
        return output, output_bias

    def extra_repr(self) -> str:
        s = f"in_features={self.input_size}"
        s += f", output_features={self.output_size_per_partition}"
        s += f", bias={self.bias is not None}"
        s += f", tp_size={get_tensor_model_parallel_world_size()}"
        s += f", gather_output={self.gather_output}"
        return s


class MergedColumnParallelLinear(ColumnParallelLinear):
    """Packed linear layers with column parallelism.

    Similar to ColumnParallelLinear, but the weight matrix is concatenated
    along the output dimension. When the weight matrix is loaded, the
    different partitions are sharded separately.

    Args:
        input_size: input dimension of the linear layer.
        output_sizes: list of output dimensions of the linear layer.
        bias: If true, add bias.
        gather_output: If true, call all-gather on output and make the output
                       available to all GPUs, otherwise, every GPU will have
                       its own output.
        skip_bias_add: This was added to enable performance optimizations where
                       bias can be fused with other element-wise operations. we
                       skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
        quant_config: Quantization configure.
    """

    def __init__(self,
                 input_size: int,
                 output_sizes: List[int],
                 bias: bool = True,
                 gather_output: bool = False,
                 skip_bias_add: bool = False,
                 params_dtype: Optional[torch.dtype] = None,
                 quant_config: Optional[QuantizationConfig] = None):
        self.output_sizes = output_sizes
        tp_size = get_tensor_model_parallel_world_size()
        assert all(output_size % tp_size == 0 for output_size in output_sizes)
        super().__init__(input_size=input_size,
                         output_size=sum(output_sizes),
                         bias=bias,
                         gather_output=gather_output,
                         skip_bias_add=skip_bias_add,
                         params_dtype=params_dtype,
                         quant_config=quant_config)

    def weight_loader(self,
                      param: Parameter,
                      loaded_weight: torch.Tensor,
                      loaded_shard_id: Optional[int] = None):

        param_data = param.data
        output_dim = getattr(param, "output_dim", None)
        # Special case for AQLM codebooks.
        is_metadata = getattr(param, "is_metadata", False)

        param_shard_splitter = getattr(param, "shard_splitter", None)

        if output_dim is not None and param_shard_splitter is not None:
            raise NotImplementedError(
                "We do not currently support output_dim != None and "
                "shard_splitter != None for a parameter. Please open an issue."
            )
        # If a parameter has defined a shard_splitter to be used for
        # the weight, it should be applied before the weight is
        # loaded/copied to the parameter. The shard_splitter applies
        # logic by using the loaded_shard_id to ensure that the loaded
        # param is loaded to the correct location
        # within the parameter defined by the linear method.
        if loaded_shard_id is None and param_shard_splitter is not None:
            raise NotImplementedError(
                "We do not currently support loaded_shard_id == None and "
                "shard_splitter != None for a parameter. Please open an issue."
            )

        # Special case for Fp8 scales.
        fp8_scales_shard_indexer = getattr(param, "fp8_scales_shard_indexer",
                                           None)

        if loaded_shard_id is None:
            # Loaded weight is already packed.
            if output_dim is None:
                temp = loaded_weight.repeat(param_data.shape)
                assert param_data.shape == temp.shape
                param_data.copy_(temp)
                return
            current_shard_offset = 0
            shard_offsets = []
            for i, output_size in enumerate(self.output_sizes):
                shard_offsets.append((i, current_shard_offset, output_size))
                current_shard_offset += output_size
            packed_dim = getattr(param, "packed_dim", None)
            for shard_id, shard_offset, shard_size in shard_offsets:
                # Special case for Quantization.
                # If quantized, we need to adjust the offset and size to account
                # for the packing.
                if packed_dim == output_dim:
                    shard_size = shard_size // param.pack_factor
                    shard_offset = shard_offset // param.pack_factor
                    # Special case for Marlin.
                    shard_size, shard_offset = adjust_marlin_shard(
                        param, shard_size, shard_offset)

                loaded_weight_shard = loaded_weight.narrow(
                    output_dim, shard_offset, shard_size)
                self.weight_loader(param, loaded_weight_shard, shard_id)
            return

        assert loaded_shard_id < len(self.output_sizes)
        tp_rank = get_tensor_model_parallel_rank()
        tp_size = get_tensor_model_parallel_world_size()
        if output_dim is not None:
            shard_offset = sum(self.output_sizes[:loaded_shard_id]) // tp_size
            shard_size = self.output_sizes[loaded_shard_id] // tp_size
            # Special case for quantization.
            # If quantized, we need to adjust the offset and size to account
            # for the packing.
            packed_dim = getattr(param, "packed_dim", None)
            if packed_dim == output_dim:
                shard_size = shard_size // param.pack_factor
                shard_offset = shard_offset // param.pack_factor
                # Special case for Marlin.
                shard_size, shard_offset = adjust_marlin_shard(
                    param, shard_size, shard_offset)

            use_bitsandbytes = getattr(param, "use_bitsandbytes", False)
            if use_bitsandbytes:
                shard_size = loaded_weight.shape[output_dim]
                shard_offset = loaded_weight.shape[output_dim] * \
                    loaded_shard_id

            param_data = param_data.narrow(output_dim, shard_offset,
                                           shard_size)
            start_idx = tp_rank * shard_size
            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                 shard_size)
        # Special case for AQLM codebooks.
        elif is_metadata:
            # metadata indicates fixed size concatenated along dim 0
            shard_size = loaded_weight.shape[0]
            shard_offset = loaded_shard_id * shard_size
            param_data = param_data.narrow(0, shard_offset, shard_size)

        # If a param_shard_splitter is defined by the LinearMethod, use it.
        elif param_shard_splitter is not None:
            logical_widths = getattr(param, "logical_widths", None)
            param_data, loaded_weight = param_shard_splitter(
                param_data, loaded_weight, loaded_shard_id, logical_widths)

        # Special case for Fp8 scales.
        elif fp8_scales_shard_indexer is not None:
            param_data, loaded_weight = fp8_scales_shard_indexer(
                param_data, loaded_weight, loaded_shard_id)

        else:
            ignore_warning = getattr(param, "ignore_warning", False)
            if not ignore_warning:
                logger.warning(
                    "Loading a weight without `output_dim` attribute in "
                    "MergedColumnParallelLinear, assume the weight is "
                    "the same for all partitions.")

        if fp8_scales_shard_indexer is None:
            if len(param_data.shape) == 0:
                param_data = param_data.reshape(1)

            if len(loaded_weight.shape) == 0:
                loaded_weight = loaded_weight.reshape(1)

        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)


class QKVParallelLinear(ColumnParallelLinear):
    """Linear layers for the attention's QKV transformation.

    Linear layers for the linear transformation of the query, key, and value
    vectors in the attention layer. The weight matrix is concatenated along
    the output dimension. The layer is parallelized along the head dimension.
    When the number of key/value heads is smaller than the number of query
    heads (e.g., multi-query/grouped-query attention), the key/value head may
    be replicated while the query heads are partitioned.

    Args:
        hidden_size: input hidden state size of the transformer.
        head_size: size of each attention head.
        total_num_heads: total number of attention query heads.
        total_num_kv_heads: total number of attention key/value heads. If
                            None, assume total_num_kv_heads = total_num_heads.
        bias: If true, add bias.
        skip_bias_add: This was added to enable performance optimizations where
                       bias can be fused with other element-wise operations. we
                       skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
        quant_config: Quantization configure.
    """

    def __init__(self,
                 hidden_size: int,
                 head_size: int,
                 total_num_heads: int,
                 total_num_kv_heads: Optional[int] = None,
                 bias: bool = True,
                 skip_bias_add: bool = False,
                 params_dtype: Optional[torch.dtype] = None,
                 quant_config: Optional[QuantizationConfig] = None):
        self.hidden_size = hidden_size
        self.head_size = head_size
        self.total_num_heads = total_num_heads
        if total_num_kv_heads is None:
            total_num_kv_heads = total_num_heads
        self.total_num_kv_heads = total_num_kv_heads
        # Divide the weight matrix along the last dimension.
        tp_size = get_tensor_model_parallel_world_size()
        self.num_heads = divide(self.total_num_heads, tp_size)
        if tp_size >= self.total_num_kv_heads:
            self.num_kv_heads = 1
            self.num_kv_head_replicas = divide(tp_size,
                                               self.total_num_kv_heads)
        else:
            self.num_kv_heads = divide(self.total_num_kv_heads, tp_size)
            self.num_kv_head_replicas = 1
        input_size = self.hidden_size
        output_size = (self.num_heads +
                       2 * self.num_kv_heads) * tp_size * self.head_size
        self.output_sizes = [
            self.num_heads * self.head_size * tp_size,  # q_proj
            self.num_kv_heads * self.head_size * tp_size,  # k_proj
            self.num_kv_heads * self.head_size * tp_size,  # v_proj 
        ]

        super().__init__(input_size=input_size,
                         output_size=output_size,
                         bias=bias,
                         gather_output=False,
                         skip_bias_add=skip_bias_add,
                         params_dtype=params_dtype,
                         quant_config=quant_config)

    def weight_loader(self,
                      param: Parameter,
                      loaded_weight: torch.Tensor,
                      loaded_shard_id: Optional[str] = None):
        param_data = param.data
        output_dim = getattr(param, "output_dim", None)
        # Special case for AQLM codebooks.
        is_metadata = getattr(param, "is_metadata", False)

        param_shard_splitter = getattr(param, "shard_splitter", None)

        if output_dim is not None and param_shard_splitter is not None:
            raise NotImplementedError(
                "We do not currently support output_dim != None and "
                "shard_splitter != None for a parameter. Please open an issue."
            )
        # If a parameter has defined a shard_splitter to be used for
        # the weight, it should be applied before the weight is
        # loaded/copied to the parameter. The shard_splitter applies
        # logic by using the loaded_shard_id to ensure that the loaded
        # param is loaded to the correct location
        # within the parameter defined by the linear method.
        if loaded_shard_id is None and param_shard_splitter is not None:
            raise NotImplementedError(
                "We do not currently support loaded_shard_id == None and "
                "shard_splitter != None for a parameter. Please open an issue."
            )

        # Special case for Fp8 scales.
        fp8_scales_shard_indexer = getattr(param, "fp8_scales_shard_indexer",
                                           None)

        if loaded_shard_id is None:
            # Loaded weight is already packed.
            if output_dim is None:
                temp = loaded_weight.repeat(param_data.shape)
                assert param_data.shape == temp.shape
                param_data.copy_(temp)
                return
            shard_offsets = [
                # (shard_id, shard_offset, shard_size)
                ("q", 0, self.total_num_heads * self.head_size),
                ("k", self.total_num_heads * self.head_size,
                 self.total_num_kv_heads * self.head_size),
                ("v", (self.total_num_heads + self.total_num_kv_heads) *
                 self.head_size, self.total_num_kv_heads * self.head_size),
            ]
            packed_dim = getattr(param, "packed_dim", None)
            for shard_id, shard_offset, shard_size in shard_offsets:
                # Special case for Quantized Weights.
                # If quantized, we need to adjust the offset and size to account
                # for the packing.
                if packed_dim == output_dim:
                    shard_size = shard_size // param.pack_factor
                    shard_offset = shard_offset // param.pack_factor

                    # Special case for Marlin.
                    shard_size, shard_offset = adjust_marlin_shard(
                        param, shard_size, shard_offset)

                loaded_weight_shard = loaded_weight.narrow(
                    output_dim, shard_offset, shard_size)
                self.weight_loader(param, loaded_weight_shard, shard_id)
            return

        tp_rank = get_tensor_model_parallel_rank()
        assert loaded_shard_id in ["q", "k", "v"]

        # If output dim is defined, use the default loading process.
        if output_dim is not None:
            if loaded_shard_id == "q":
                shard_offset = 0
                shard_size = self.num_heads * self.head_size
            elif loaded_shard_id == "k":
                shard_offset = self.num_heads * self.head_size
                shard_size = self.num_kv_heads * self.head_size
            elif loaded_shard_id == "v":
                shard_offset = (self.num_heads +
                                self.num_kv_heads) * self.head_size
                shard_size = self.num_kv_heads * self.head_size
            # Special case for Quantized Weights.
            # If quantized, we need to adjust the offset and size to account
            # for the packing.
            packed_dim = getattr(param, "packed_dim", None)
            if packed_dim == output_dim:
                shard_size = shard_size // param.pack_factor
                shard_offset = shard_offset // param.pack_factor

                # Special case for Marlin.
                shard_size, shard_offset = adjust_marlin_shard(
                    param, shard_size, shard_offset)

            use_bitsandbytes = getattr(param, "use_bitsandbytes", False)
            if use_bitsandbytes:
                orig_qkv_offsets = {
                    "q": (0, self.num_heads * self.head_size),
                    "k": (self.num_heads * self.head_size,
                          self.num_kv_heads * self.head_size),
                    "v":
                    ((self.num_heads + self.num_kv_heads) * self.head_size,
                     self.num_kv_heads * self.head_size),
                    "total":
                    ((self.num_heads + 2 * self.num_kv_heads) * self.head_size,
                     0)
                }
                shard_size, shard_offset = adjust_bitsandbytes_shard(
                    param, orig_qkv_offsets, loaded_shard_id)

            param_data = param_data.narrow(output_dim, shard_offset,
                                           shard_size)
            if loaded_shard_id == "q":
                shard_id = tp_rank
            else:
                shard_id = tp_rank // self.num_kv_head_replicas
            start_idx = shard_id * shard_size
            loaded_weight = loaded_weight.narrow(output_dim, start_idx,
                                                 shard_size)
        # Special case for for AQLM codebooks.
        elif is_metadata:
            # metadata indicates fixed size concatenated along dim 0
            shard_size = loaded_weight.shape[0]
            shard_index = ["q", "k", "v"].index(loaded_shard_id)
            param_data = param_data.narrow(0, shard_index * shard_size,
                                           shard_size)
        # If a param_shard_splitter is defined by the LinearMethod, use it.
        elif param_shard_splitter is not None:
            logical_widths = getattr(param, "logical_widths", None)
            param_data, loaded_weight = param_shard_splitter(
                param_data, loaded_weight, loaded_shard_id, logical_widths)

        # Special case for Fp8 scales.
        elif fp8_scales_shard_indexer is not None:
            param_data, loaded_weight = fp8_scales_shard_indexer(
                param_data, loaded_weight, loaded_shard_id)
        else:
            ignore_warning = getattr(param, "ignore_warning", False)
            if not ignore_warning:
                logger.warning(
                    "Loading a weight without `output_dim` attribute in "
                    "QKVParallelLinear, assume the weight is the same "
                    "for all partitions.")

        if len(param_data.shape) == 0:
            param_data = param_data.reshape(1)

        if len(loaded_weight.shape) == 0:
            loaded_weight = loaded_weight.reshape(1)

        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)


class RowParallelLinear(LinearBase):
    """Linear layer with row parallelism.

    The linear layer is defined as Y = XA + b. A is parallelized along
    its first dimension and X along its second dimension as:
               -   -
              | A_1 |
              | .   |
          A = | .   |        X = [X_1, ..., X_p]
              | .   |
              | A_p |
               -   -
    Arguments:
        input_size: first dimension of matrix A.
        output_size: second dimension of matrix A.
        bias: If true, add bias. Note that bias is not parallelized.
        input_is_parallel: If true, we assume that the input is already
                           split across the GPUs and we do not split
                           again.
        skip_bias_add: This was added to enable performance optimization where
                       bias can be fused with other element-wise operations.
                       We skip adding bias but instead return it.
        params_dtype: Data type for the parameters.
        quant_config: Quantization configure.
    """

    def __init__(self,
                 input_size: int,
                 output_size: int,
                 bias: bool = True,
                 input_is_parallel: bool = True,
                 skip_bias_add: bool = False,
                 params_dtype: Optional[torch.dtype] = None,
                 reduce_results: bool = True,
                 quant_config: Optional[QuantizationConfig] = None):
        super().__init__(input_size, output_size, skip_bias_add, params_dtype,
                         quant_config)

        self.input_is_parallel = input_is_parallel
        self.reduce_results = reduce_results

        # Divide the weight matrix along the last dimension.
        self.tp_size = get_tensor_model_parallel_world_size()
        self.input_size_per_partition = divide(input_size, self.tp_size)
        assert self.quant_method is not None
        self.quant_method.create_weights(
            layer=self,
            input_size_per_partition=self.input_size_per_partition,
            output_partition_sizes=[self.output_size],
            input_size=self.input_size,
            output_size=self.output_size,
            params_dtype=self.params_dtype,
            weight_loader=self.weight_loader)
        if not reduce_results and (bias and not skip_bias_add):
            raise ValueError("When not reduce the results, adding bias to the "
                             "results can lead to incorrect results")

        if bias:
            self.bias = Parameter(
                torch.empty(self.output_size, dtype=params_dtype))
            set_weight_attrs(self.bias, {
                "output_dim": 0,
                "weight_loader": self.weight_loader,
            })
        else:
            self.register_parameter("bias", None)

    def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
        # Special case for Fp8 scales.
        fp8_scales_shard_indexer = getattr(param, "fp8_scales_shard_indexer",
                                           None)

        tp_rank = get_tensor_model_parallel_rank()
        input_dim = getattr(param, "input_dim", None)
        param_data = param.data
        if input_dim is not None:
            shard_size = param_data.shape[input_dim]
            start_idx = tp_rank * shard_size
            loaded_weight = loaded_weight.narrow(input_dim, start_idx,
                                                 shard_size)

        # Special case for Fp8 scales.
        elif fp8_scales_shard_indexer is not None:
            param_data, loaded_weight = fp8_scales_shard_indexer(param_data,
                                                                 loaded_weight,
                                                                 shard_id=0)

        if fp8_scales_shard_indexer is None and len(loaded_weight.shape) == 0:
            loaded_weight = loaded_weight.reshape(1)

        assert param_data.shape == loaded_weight.shape
        param_data.copy_(loaded_weight)

    def forward(self, input_):
        # Set up backprop all-reduce.
        if self.input_is_parallel:
            input_parallel = input_
        else:
            tp_rank = get_tensor_model_parallel_rank()
            splitted_input = split_tensor_along_last_dim(
                input_, num_partitions=self.tp_size)
            input_parallel = splitted_input[tp_rank].contiguous()

        # Matrix multiply.
        assert self.quant_method is not None
        output_parallel = self.quant_method.apply(self, input_parallel)
        if self.reduce_results and self.tp_size > 1:
            output_ = tensor_model_parallel_all_reduce(output_parallel)
        else:
            output_ = output_parallel

        if not self.skip_bias_add:
            output = output_ + self.bias if self.bias is not None else output_
            output_bias = None
        else:
            output = output_
            output_bias = self.bias
        return output, output_bias

    def extra_repr(self) -> str:
        s = f"input_features={self.input_size_per_partition}"
        s += f", output_features={self.output_size}"
        s += f", bias={self.bias is not None}"
        s += f", tp_size={self.tp_size}"
        s += f", reduce_results={self.reduce_results}"
        return s
```


## Evaluation

Evaluated on the Open LLM Leaderboard evaluations through vLLM.

### Open LLM Leaderboard evaluation scores
|                      | Phi-3-mini-128k-instruct-FP8       | neuralmagic/Phi-3-mini-128k-instruct-FP8<br>(this model) |
| :------------------: | :----------------------: | :------------------------------------------------: |
| arc-c<br>25-shot     | 63.65                    | 63.31                                              |
| hellaswag<br>10-shot | 79.76                    | 79.44                                              |
| mmlu<br>5-shot       | 68.10                    | 68.08                                              |
| truthfulqa<br>0-shot | 53.97                    | 53.76                                              |
| winogrande<br>5-shot | 73.72                    | 72.45                                              |
| gsm8k<br>5-shot      | 75.59                    | 72.86                                              |
| **Average<br>Accuracy**  | **69.13**                    |              **68.32**                                      |
| **Recovery**             | **100%**                     |              **98.82%**                                     |